[libcxx] Add atomic_support.h header to src that handles needed atomic operations.

[EricWF]

Summary

In some places in libc++ we need to use the __atomic_* builtins. This patch adds a header that provides access to those builtins in a uniform way from within the dylib source.

If the compiler building the dylib does not support these builtins then a warning is issued.

Only relaxed loads are needed within the headers. A singe function to do these relaxed loads has been added to <memory>.

This patch applies the new atomic builtins to __shared_count and call_once.

[jroelofs]

include/__atomic

85 ↗

(On Diff #27557)

Should these fall back on the __sync_* ones instead of just dropping the atomicness? It seems a little precarious to silently default to non-atomic implementations for these (unless this is for a single-threaded build of the library).

[EricWF]

Respond to @jroelofs comments.

include/__atomic

85 ↗

(On Diff #27557)

Yes and no. The only atomic operations needed in the headers are relaxed atomic loads. AFAIK all loads are relaxed atomic loads on x86 so not using atomic builtins isn't a big deal (its what we already do).

The rest of the atomic operations happen within the dylib. I agree we don't want to drop the atomicness on these unless it is a single threaded build.

What we really should do is limit the libc++ dylib build to compilers that provide the required atomic intrinsics (single-threaded builds not withstanding). If the compiler is not supported then the build should fail. For GCC and Clang the required versions would be 4.7 and 3.3 (possibly lower) respectively.

Since most of the stuff in __atomic is only used in the dylib build perhaps it should be moved out of the shipped headers. I'll figure out how to handle that tomorrow.

[jroelofs]

include/__atomic

85 ↗

(On Diff #27557)

AFAIK all loads are relaxed atomic loads on x86 so not using atomic builtins isn't a big deal

What about PPC-darwin?

What we really should do is limit the libc++ dylib build to compilers that provide the required atomic intrinsics (single-threaded builds not withstanding)

Agreed.

[majnemer]

src/mutex.cpp

255–267

Can you please leave a note that says that these relaxed stores pair with the relaxed load in the header? The atomic nature of the stores look superfluous because they are performed under the mutex.

[EricWF]

Remove the __atomic header and put the contents in the source files. Enforce that we have atomics when building libc++.

[EricWF]

Address @majnemer's request for a comment.

[mclow.lists]

I really like that this stuff (support/atomic) is only used in the dylib.

Does this patch fix: https://llvm.org/bugs/show_bug.cgi?id=22803 ?

And do we have an outstanding bug against call_once ?

[EricWF]

In D10406#192273, @mclow.lists wrote:

I really like that this stuff (support/atomic) is only used in the dylib.

Does this patch fix: https://llvm.org/bugs/show_bug.cgi?id=22803 ?

Same patch. (It links here :P)

And do we have an outstanding bug against call_once ?

No outstanding bug. The call_once changes are to fix/suppress a couple of TSAN diagnostics. (See the TSAN buildbot)

[EricWF]

In D10406#192279, @EricWF wrote:

In D10406#192273, @mclow.lists wrote:

Does this patch fix: https://llvm.org/bugs/show_bug.cgi?id=22803 ?

Woops Misread that comment. This patch does not currently fix that bug but it can and will once I update it. The change is trivial.

[EricWF]

1. Fix increment/decrement consistency to fix PR22803.
2. Put functions in atomic_support.h inside an inline namespace.
3. Generate a build warning on MSVC when atomics aren't supported.
4. Fix bad assertion inside shared_ptr race condition test case.

I don't think we should make any attempt to support GCC and Clang versions that do not support the __atomic* builtins. This would limit *building* (not using) libc++ to GCC 4.7 and beyond and clang 3.3 (possible earlier). I don't have an issue with this.

[EricWF]

Remove unrelated change in <thread>.

[mclow.lists]

Except for the stuff in <thread>, this looks good.

include/thread

338 ↗

(On Diff #29126)

I think this snuck in from somewhere else. Please don't commit this.

include/memory

 
 _LIBCPP_BEGIN_NAMESPACE_STD
 
+template <class _ValueType>
+inline _LIBCPP_ALWAYS_INLINE
+_ValueType __libcpp_relaxed_load(_ValueType const* __value) {
+#if !defined(_LIBCPP_HAS_NO_THREADS) && \
+    defined(__ATOMIC_RELAXED) &&        \
+    (__has_builtin(__atomic_load_n) || _GNUC_VER >= 407)
+    return __atomic_load_n(__value, __ATOMIC_RELAXED);
+#else
+    return *__value;
+#endif
+}
+
 // addressof moved to <__functional_base>
 
 template <class _Tp> class allocator;

     void __add_shared() _NOEXCEPT;
     bool __release_shared() _NOEXCEPT;
     _LIBCPP_INLINE_VISIBILITY
-    long use_count() const _NOEXCEPT {return __shared_owners_ + 1;}
+    long use_count() const _NOEXCEPT {
+        return __libcpp_relaxed_load(&__shared_owners_) + 1;
+    }
 };
 
 class _LIBCPP_TYPE_VIS __shared_weak_count

include/mutex

 #include <__config>
 #include <__mutex_base>
 #include <functional>
+#include <memory>
 #ifndef _LIBCPP_HAS_NO_VARIADICS
 #include <tuple>
 #endif

 void
 call_once(once_flag& __flag, _Callable&& __func, _Args&&... __args)
 {
-    if (__flag.__state_ != ~0ul)
+    if (__libcpp_relaxed_load(&__flag.__state_) != ~0ul)
     {
         typedef tuple<typename decay<_Callable>::type, typename decay<_Args>::type...> _Gp;
         __call_once_param<_Gp> __p(_Gp(__decay_copy(_VSTD::forward<_Callable>(__func)),

 void
 call_once(once_flag& __flag, _Callable __func)
 {
-    if (__flag.__state_ != ~0ul)
+    if (__libcpp_relaxed_load(&__flag.__state_) != ~0ul)
     {
         __call_once_param<_Callable> __p(__func);
         __call_once(__flag.__state_, &__p, &__call_once_proxy<_Callable>);

src/memory.cpp

 #include "mutex"
 #include "thread"
 #endif
+#include "support/atomic_support.h"
 
 _LIBCPP_BEGIN_NAMESPACE_STD
 

 inline T
 increment(T& t) _NOEXCEPT
 {
-    return __sync_add_and_fetch(&t, 1);
+    return __libcpp_atomic_add(&t, 1);
 }
 
 template <class T>
 inline T
 decrement(T& t) _NOEXCEPT
 {
-    return __sync_add_and_fetch(&t, -1);
+    return __libcpp_atomic_add(&t, -1);
 }
 
 }  // namespace

 __shared_weak_count*
 __shared_weak_count::lock() _NOEXCEPT
 {
-    long object_owners = __shared_owners_;
+    long object_owners = __libcpp_atomic_load(&__shared_owners_);
     while (object_owners != -1)
     {
-        if (__sync_bool_compare_and_swap(&__shared_owners_,
-                                         object_owners,
+        if (__libcpp_atomic_compare_exchange(&__shared_owners_,
+                                             &object_owners,
                                              object_owners+1))
             return this;
-        object_owners = __shared_owners_;
     }
     return 0;
 }

src/mutex.cpp

 #include "limits"
 #include "system_error"
 #include "cassert"
+#include "support/atomic_support.h"
 
 _LIBCPP_BEGIN_NAMESPACE_STD
 #ifndef _LIBCPP_HAS_NO_THREADS

 static pthread_cond_t  cv  = PTHREAD_COND_INITIALIZER;
 #endif
 
+/// NOTE: Changes to flag are done via relaxed atomic stores
+///       even though the accesses are protected by a mutex because threads
+///       just entering 'call_once` concurrently read from flag.
 void
 __call_once(volatile unsigned long& flag, void* arg, void(*func)(void*))
 {

         try
         {
 #endif  // _LIBCPP_NO_EXCEPTIONS
-            flag = 1;
+            __libcpp_relaxed_store(&flag, 1ul);
             pthread_mutex_unlock(&mut);
             func(arg);
             pthread_mutex_lock(&mut);
-            flag = ~0ul;
+            __libcpp_relaxed_store(&flag, ~0ul);
             pthread_mutex_unlock(&mut);
             pthread_cond_broadcast(&cv);
 #ifndef _LIBCPP_NO_EXCEPTIONS

         catch (...)
         {
             pthread_mutex_lock(&mut);
-            flag = 0ul;
+            __libcpp_relaxed_store(&flag, 0ul);
             pthread_mutex_unlock(&mut);
             pthread_cond_broadcast(&cv);
             throw;

src/support/atomic_support.h

+#ifndef ATOMIC_SUPPORT_H
+#define ATOMIC_SUPPORT_H
+
+#include "__config"
+#include "memory" // for __libcpp_relaxed_load
+
+#if defined(__clang__) && __has_builtin(__atomic_load_n)             \
+                       && __has_builtin(__atomic_store_n)            \
+                       && __has_builtin(__atomic_add_fetch)          \
+                       && __has_builtin(__atomic_compare_exchange_n) \
+                       && defined(__ATOMIC_RELAXED)                  \
+                       && defined(__ATOMIC_CONSUME)                  \
+                       && defined(__ATOMIC_ACQUIRE)                  \
+                       && defined(__ATOMIC_RELEASE)                  \
+                       && defined(__ATOMIC_ACQ_REL)                  \
+                       && defined(__ATOMIC_SEQ_CST)
+#   define _LIBCPP_HAS_ATOMIC_BUILTINS
+#elif !defined(__clang__) && defined(_GNUC_VER) && _GNUC_VER >= 407
+#   define _LIBCPP_HAS_ATOMIC_BUILTINS
+#endif
+
+#if !defined(_LIBCPP_HAS_ATOMIC_BUILTINS) && !defined(_LIBCPP_HAS_NO_THREADS)
+#warning Building libc++ without __atomic builtins is unsupported
+#endif
+
+_LIBCPP_BEGIN_NAMESPACE_STD
+
+#if defined(_LIBCPP_HAS_ATOMIC_BUILTINS) && !defined(_LIBCPP_HAS_NO_THREADS)
+
+enum __libcpp_atomic_order {
+    _AO_Relaxed = __ATOMIC_RELAXED,
+    _AO_Consume = __ATOMIC_CONSUME,
+    _AO_Aquire  = __ATOMIC_ACQUIRE,
+    _AO_Release = __ATOMIC_RELEASE,
+    _AO_Acq_Rel = __ATOMIC_ACQ_REL,
+    _AO_Seq     = __ATOMIC_SEQ_CST
+};
+
+template <class _ValueType, class _FromType>
+inline _LIBCPP_INLINE_VISIBILITY
+void __libcpp_atomic_store(_ValueType* __dest, _FromType __val,
+                           int __order = _AO_Seq)
+{
+    __atomic_store_n(__dest, __val, __order);
+}
+
+template <class _ValueType, class _FromType>
+inline _LIBCPP_INLINE_VISIBILITY
+void __libcpp_relaxed_store(_ValueType* __dest, _FromType __val)
+{
+    __atomic_store_n(__dest, __val, _AO_Relaxed);
+}
+
+template <class _ValueType>
+inline _LIBCPP_INLINE_VISIBILITY
+_ValueType __libcpp_atomic_load(_ValueType const* __val,
+                                int __order = _AO_Seq)
+{
+    return __atomic_load_n(__val, __order);
+}
+
+template <class _ValueType, class _AddType>
+inline _LIBCPP_INLINE_VISIBILITY
+_ValueType __libcpp_atomic_add(_ValueType* __val, _AddType __a,
+                               int __order = _AO_Seq)
+{
+    return __atomic_add_fetch(__val, __a, __order);
+}
+
+template <class _ValueType>
+inline _LIBCPP_INLINE_VISIBILITY
+bool __libcpp_atomic_compare_exchange(_ValueType* __val,
+    _ValueType* __expected, _ValueType __after,
+    int __success_order = _AO_Seq,
+    int __fail_order = _AO_Seq)
+{
+    return __atomic_compare_exchange_n(__val, __expected, __after, true,
+                                       __success_order, __fail_order);
+}
+
+#else // _LIBCPP_HAS_NO_THREADS
+
+enum __libcpp_atomic_order {
+    _AO_Relaxed,
+    _AO_Consume,
+    _AO_Acquire,
+    _AO_Release,
+    _AO_Acq_Rel,
+    _AO_Seq
+};
+
+template <class _ValueType, class _FromType>
+inline _LIBCPP_INLINE_VISIBILITY
+void __libcpp_atomic_store(_ValueType* __dest, _FromType __val,
+                           int = 0)
+{
+    *__dest = __val;
+}
+
+template <class _ValueType>
+inline _LIBCPP_INLINE_VISIBILITY
+_ValueType __libcpp_atomic_load(_ValueType const* __val,
+                                int = 0)
+{
+    return *__val;
+}
+
+template <class _ValueType, class _AddType>
+inline _LIBCPP_INLINE_VISIBILITY
+_ValueType __libcpp_atomic_add(_ValueType* __val, _AddType __a,
+                               int = 0)
+{
+    return *__val += __a;
+}
+
+template <class _ValueType>
+inline _LIBCPP_INLINE_VISIBILITY
+bool __libcpp_atomic_compare_exchange(_ValueType* __val,
+    _ValueType* __expected, _ValueType __after,
+    int = 0, int = 0)
+{
+    if (*__val == *__expected) {
+        *__val = __after;
+        return true;
+    }
+    *__expected = *__val;
+    return false;
+}
+
+#endif // _LIBCPP_HAS_NO_THREADS
+
+_LIBCPP_END_NAMESPACE_STD
+
+#endif // ATOMIC_SUPPORT_H

test/libcxx/utilities/memory/util.smartptr/race_condition.pass.cpp

+//===----------------------------------------------------------------------===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is dual licensed under the MIT and the University of Illinois Open
+// Source Licenses. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+// UNSUPPORTED: libcpp-has-no-threads
+//
+// <memory>
+//
+// class shared_ptr
+//
+// This test attempts to create a race condition surrounding use_count()
+// with the hope that TSAN will diagnose it.
+
+#include <memory>
+#include <atomic>
+#include <thread>
+#include <cassert>
+
+typedef std::shared_ptr<int> Ptr;
+typedef std::weak_ptr<int> WeakPtr;
+
+std::atomic_bool KeepRunning;
+
+struct TestRunner {
+    TestRunner(Ptr xx) : x(xx) {}
+    void operator()() {
+        while (KeepRunning) {
+            // loop to prevent always checking the atomic.
+            for (int i=0; i < 100000; ++i) {
+                Ptr x2 = x; // increment shared count
+                WeakPtr x3 = x; // increment weak count
+                Ptr x4 = x3.lock(); // increment shared count via lock
+                WeakPtr x5 = x3; // increment weak count
+            }
+        }
+    }
+    Ptr x;
+};
+
+void run_test(Ptr p) {
+    // Start the test thread.
+    KeepRunning = true;
+    TestRunner r(p);
+    std::thread t1(r);
+    assert(p.use_count() == 3);
+    // Run until we witness 25 use count changes via both
+    // shared and weak pointer methods.
+    WeakPtr w = p;
+    int shared_changes_count = 0;
+    int weak_changes_count = 0;
+    while (shared_changes_count < 25 && weak_changes_count < 25) {
+        // check use_count on the shared_ptr
+       int last = p.use_count();
+       int new_val = p.use_count();
+       assert(last >= 3);
+       assert(new_val >= 3);
+       if (last != new_val) ++shared_changes_count;
+       // Check use_count on the weak_ptr
+       last = w.use_count();
+       new_val = w.use_count();
+       assert(last >= 3);
+       assert(new_val >= 3);
+       if (last != new_val) ++weak_changes_count;
+    }
+    // kill the test thread.
+    KeepRunning = false;
+    t1.join();
+    assert(p.use_count() == 3);
+}
+
+int main() {
+  {
+    // Test with out-of-place shared_count.
+    Ptr p(new int(42));
+    run_test(p);
+    assert(p.use_count() == 1);
+  }
+  {
+    // Test with in-place shared_count.
+    Ptr p = std::make_shared<int>(42);
+    run_test(p);
+    assert(p.use_count() == 1);
+  }
+}